Beat Tracking using Recurrent Neural Network: A Transfer Learning Approach
Davide Fiocchi, Michele Buccoli, Massimiliano Zanoni, Fabio Antonacci, Augusto Sarti · 2018
Deep learning networks have been successfully applied to solve a large number of tasks. The effectiveness of deep learning networks is limited by the amount and the variety of data used for the training. For this reason, deep-learning networks can be applied in scenarios where a huge amount of data are available. In music information retrieval, this is the case of popular genres due to the wider availability of annotated music pieces. Instead, to find sufficient and useful data is a hard task for non widespread genres, like, for instance, traditional and folk music. To address this issue, Transfer Learning has been proposed, i.e., to train a network using a large available dataset and then transfer the learned knowledge (the hierarchical representation) to another task. In this work, we propose an approach to apply transfer learning for beat tracking. We use a deep BLSTM-based RNN as the starting network trained on popular music, and we transfer it to track beats of Greek folk music. In order to evaluate the effectiveness of our approach, we collect a dataset of Greek folk music, and we manually annotate the pieces.